Prediction of lymph node metastasis in cN0 papillary thyroid carcinoma using a random forest–based radiomics model integrating contrast-enhanced ultrasound and clinical data

Objectives The aim of this study was to evaluate a radiomics model integrating contrast-enhanced ultrasound (CEUS) features and clinical indicators for predicting lymph node metastasis (LNM) in patients with papillary thyroid carcinoma (PTC) classified as clinically node-negative (cN0). Methods A total of 604 patients with preoperative cN0 PTC who underwent routine ultrasound and CEUS examinations were enrolled. Radiomics features were extracted from ultrasound images and integrated with contrast parameters and clinical indicators to construct machine learning models. The diagnostic performance of different models in predicting central lymph node metastasis (CLNM) and overall LNM was compared. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression analysis. Performance metrics were compared to determine the optimal model. Receiver operating characteristic (ROC) curves, precision–recall (PR) curves, calibration curves, and decision curves were generated to illustrate differences in diagnostic performance. SHapley Additive exPlanations (SHAP) were applied for model interpretability. Results Data were divided into five groups. The Clinical + Radiomics + Contrast parameter random forest (RF) model demonstrated the highest predictive performance for both CLNM and LNM. The two models incorporated different predictive indicators. The area under the curve (AUC) values were 0.928 and 0.942, respectively. SHAP visualization identified 20 important features for each optimal model. Radiomics features, particularly those derived from CEUS images, along with contrast parameters, demonstrated significant predictive value in addition to clinical characteristics. Conclusions Radiomics features and CEUS contrast parameters enhanced the diagnostic efficacy of predictive models. The RF model integrating clinical, radiomics, and contrast parameter data demonstrated the highest predictive accuracy for CLNM and LNM in patients with cN0 PTC.